Use Cosyte with AI agents

An AI coding agent can work with the Cosyte packages in three ways: it can call our parsers through the MCP server in @cosyte/cli, it can load our Agent Skills, and it can read this site's documentation as plain Markdown instead of HTML.

The MCP server in @cosyte/cli​

@cosyte/cli ships an MCP (Model Context Protocol) server, cosyte-mcp. It is a local stdio subprocess that your MCP client starts on your machine, not a hosted endpoint.

Register it​

Add this to your MCP client's configuration:

{
  "mcpServers": {
    "cosyte": { "command": "npx", "args": ["-y", "--package", "@cosyte/cli", "cosyte-mcp"] }
  }
}

--package is required. The shorter ["-y", "@cosyte/cli", "mcp"] does not work: it fails with could not determine executable to run.

In Claude Code, this command registers the same server:

claude mcp add --transport stdio cosyte -- npx -y --package @cosyte/cli cosyte-mcp

Everything after -- is the command that starts the server. See Anthropic's claude mcp add documentation for scopes and other options.

Versions 0.0.1 and 0.0.2 cannot be installed​

@cosyte/cli 0.0.1 and 0.0.2 are published on npm, but they cannot be installed. npm install, npx and the registration above all fail on them with an ENOENT error, and a published version cannot be changed. Use a later version.

The four tools​

Each tool runs the same code as the matching cosyte command in a terminal, so the agent and the terminal get identical results. Every tool publishes an outputSchema, and a message that parses but is invalid is a successful validate call that reports the verdict, not a tool error.

What reaches the agent​

The CLI's PHI posture carries over to the MCP server, with one door removed:

• There is no --unsafe-show-values option on the agent surface.
• A tool result carries the data you asked for: the parsed model, or the converted Bundle. Parsing a message hands its contents to the agent.
• A tool error carries only a value-free diagnostic: a stable code and a position, never a name, a date of birth, a medical record number or a field value.
• The CLI writes no temporary files and logs to no file.

The @cosyte/cli documentation covers the result contract and every command.

Agent Skills for each package​

cosyte/agent-skills (MIT) holds an Agent Skill (SKILL.md) for each @cosyte package, a Cursor rule for each skill, and an AGENTS.md. A skill tells a coding agent which package fits a task and how to use it: the install line, an example with synthetic data, the key API names, the package's current limits and known defects, and how to handle PHI.

In Claude Code, install them as a plugin from your shell:

claude plugin marketplace add cosyte/agent-skills
claude plugin install cosyte@cosyte

Inside a Claude Code session, the same two steps are /plugin marketplace add cosyte/agent-skills and /plugin install cosyte@cosyte. To copy the skill folders instead, or to use them with OpenAI Codex or Cursor, follow the repository's README. It also says what has not been tested yet: its plugin manifests have not been run through claude plugin validate, and an install has not been run in each tool.

This site as Markdown​

Every file below covers the newest released version of each package. Superseded versions are left out, and so is any package that has not published documentation yet.

The index follows the llms.txt format. It links each package's documentation and each package's full-text file, so an agent can start there and fetch only what it needs. For example, the Markdown copy of https://docs.cosyte.com/hl7/intro is https://docs.cosyte.com/hl7/intro.md.

A Markdown copy is the page's source with the front matter removed and its links made absolute, so it reads correctly on its own. In a full-text file, each page opens with a URL: line naming the page and a Markdown: line naming its copy.